1. Here's the output from Claude. The prompt I used was You're a summary expert. Review this text of a recent meeting and identify the key learnings. Present the learnings in a non-technical language so that it is easy for any AI newcomer to understand. Present an overall summary first, then each learning in bullet point format, followed by the explanation and then the example. Complete the learnings by identifying a number of action items I can take away from this meeting. 2. Good quality - able to take large volumes of data and provide a concise and accurate summary, especially in non-technical language. Good comments posted so far, although I would not know much about AI yet. 3. Success of this feature - around the continued ease of use, the continual learning of the platform, and more importantly, what business problem does it solve?

@Phil L Question to which I'm not sure - to get verified sources so as to present accurate listings, would not that mean RAG is needed? Or maybe I misunderstood, so appreciate any clarification!

@Masahiro Teramoto i think it's great to see what is needed vs what is not. Agree on LLM, RAG, and Embeddings.

PROBLEM I'M THINKING ABOUT: Wealth managers/advisors managing 100+ clients can't systematically prioritise who needs proactive outreach, causing high-value clients to churn silently and them to waste 30+ minutes daily on manual triage. Q1 - WHAT PROBLEM: Wealth managers/advisors rely on memory and generic CRM reminders, missing at-risk clients until they leave. With 100+ relationships, systematic prioritisation is impossible manually. This results in a 15-25% lower retention than firms with proactive contact systems. Q2 - SHOULD WE USE AI: Rules-based systems (contact every 365 days, flag portfolio down 15%) lack context. A 10% drop means different things for different clients. AI can reason: 'Conservative client nearing retirement + 10% drop + 60 days no contact = URGENT' while 'Aggressive client + 10% drop = expected volatility.' Talking points require understanding full client context, which AI excels at. Q3 - DO WE HAVE THE DATA: We need client data (AUM, risk profile, and tenure), portfolio data (holdings and performance), interaction history (last contact and meeting notes), and market data (benchmarks and sector returns). We should be able to obtain most from CRM/portfolio systems. An identified gap is that life events require NLP extraction from notes. MVP uses mock data mirroring real structure. Q4 - TRANSLATE TO A MODEL: Model scores all clients (0-100) using weighted factors: recency (30%), performance vs. risk profile (25%), life events (25%), churn indicators (20%). Outputs daily top 5-7 priorities with score, reason, and AI-generated talking points. Wealth managers/advisors see a one-line summary, click for details, and mark as contacted. Q5 - USER EXPERIENCE: Daily email briefing at 8 AM → advisor reviews in 3-5 min → clicks to see full context → schedules call or marks contacted. Trust is built through score explainability (show the math), override capability (dismiss/snooze), validation (prove past suggestions were valuable), and gradual rollout (observe-only → pilot → full adoption).

@Peculiar Ediomo-Abasi This is definitely a worthwhile endeavour. Some thoughts: - how do other public health bodies anticipate/deal with stockpiling essential medication? - could technology (e.g. drone delivery) be a possible workaround? - challenges of data accuracy could impact your solution - is there a similar issue in an adjacent industry (e.g. supermarkets) and how do they handle it? Just some thoughts, but keep going as it's definitely a worthwhile endeavour!

Hi everyone, My name is Jerel and I'm a Product Owner in the UK. Looking to increase my knowledge of AI and build something useful :) Fun fact: I love binge watching TV shows, most recent being Stranger Things.

@Sid Arora I would say Season 4 was the best (of the 5), but worth watching till the end for closure! :)

PRODUCT NAME: Claude WHAT IT DOES: Depending on the prompt, it answers my questions, provides write-ups, and makes recommendations based on the information it has analysed. WHAT CAN'T BE FULLY SPEC'D: The prompts may differ according to what the user needs; hence the challenge of trying to fully spec'd this at the outset. WHY AI (NOT SOMETHING SIMPLER): There is a vast amount of data that needs to be synthesized and analysed, which would take a couple of days using a multitude of tools and applications. AI has reduced the cognitive workload and time taken. HOW IT HANDLES BEING WRONG: This is where it gets tricky as it is possible that the answers given may not be accurate. But the user may or may not know that. Despite the potential for inaccuracies, the user still receives an answer so it is clear the user experience has been designed well (maybe a bit too well if the user trusts every Claude response). ONE THING THE TRADITIONAL PM WOULD MISS: Without understanding AI or the potential for errors, a traditional PM might ship this as it would be deemed as the 'user getting what they want'.

@Akshun Gulati I've hit the same problem when I use ChatGPT - while the AI tool may be quick at generating a response (in this case, an invoice), it needs human discernment to assess the accuracy of the response. I've found that if I add a follow-up prompt to challenge the original response, ChatGPT will try to calibrate. It may also take several prompts to get a respons closer to what you would expect. But yes, it feels like there's a long way to go before humans get replaced :)